Parenting & Education: Raising Kids in the AI Era

AI maxes out the speed of "finding the answer," but real learning happens during the struggle. Bypass the struggle and you bypass the learning.

Robert Bjork's "desirable difficulties": what makes learning easier often makes memory more fragile. Slamecka & Graf (1978) generation effect: generating your own answer doubles long-term retention versus reading someone else's. Bastani et al. (2024, Wharton working paper): high schoolers practicing math with GPT-4 saw a 48% accuracy boost during practice, but scored ~17% worse than controls on the post-test when AI was taken away — AI did the practice, then replaced the learning. MIT Media Lab (Kosmyna et al. 2025) used EEG to show significantly lower brain activity when writing with ChatGPT versus unaided.

Learning is not information transport — it's neural circuits forming through error → correction → consolidation. AI handing over a perfect answer is like watching someone else work out. The muscle does not grow on you.

① Total ban — kids use it anyway out of sight, never learning to judge quality. ② Trusting completed homework as evidence of learning. ③ Parents themselves running essays through AI before the teacher sees them — what the child learns is performance, not learning.

AI can be a scaffold (taken down once the structure stands) or a crutch (never put down). The difference is not the tool — it's how it's used.

Risko & Gilbert (2016, Trends in Cognitive Sciences) on cognitive offloading: outsourcing cognition to tools boosts short-term efficiency but erodes independent capacity. GPS study (Dahmani & Bohbot 2020, Scientific Reports): heavy long-term GPS users showed significantly less hippocampal gray matter and faster spatial-memory decline. Gerlich (2025, Societies) surveyed 666 AI users: frequency of AI use negatively correlated with critical-thinking scores, with metacognition as the mediator. In Vygotsky's ZPD, the defining feature of a scaffold is that it comes down. One that doesn't isn't a scaffold — it's a prosthetic.

Brains obey use-it-or-lose-it. A child who calls AI after 30 seconds of struggle is paying AI to grow the circuit for him. Homework done in the short term; capacity never built in the long term.

① Blanket ban — they use it out of sight, doubly, with no one to teach quality. ② Blanket permission — five seconds of struggle, then offload, and meta-capacity quietly disappears. ③ Parents heavily relying on AI for emails and decisions while demanding the child think unaided — kids learn what you do, not what you say.

When the marginal cost of generating an answer approaches zero, knowing what you know and what you don't becomes the core skill. Metacognition is what splits two kids using the same AI into wildly different outcomes.

Flavell (1979, American Psychologist) founded metacognition — awareness and monitoring of one's own thinking. Dunning-Kruger: the less capable, the more overconfident. Hattie & Donoghue (2016) large meta-analysis: metacognitive strategies have roughly twice the effect size of standard instructional strategies. A second finding from Bastani et al. (2024): swapping the "answer-giving GPT" for a "Socratic-prompt GPT" preserved learning — the difference being precisely whether metacognition was triggered.

AI answers always look understood. Children weak in metacognition mistake cognitive fluency for mastery and crumble on probing. Strong metacognition means self-checking — the prerequisite for "using AI well."

① Treating test scores as proof of mastery — scores measure performance, not metacognition. ② Doing all the monitoring for the child (you check homework, you flag the error) — meta-capacity stays externalized on you. ③ Accepting AI output uncritically yourself — kids learn "answer = truth."

AI can already write most everyday code. "Knows Python" is no longer the headline skill. The real value of programming education is computational thinking — decomposing complex problems into executable steps, locating causality, debugging assumptions. AI cannot replace this, because using AI well requires it.

Jeannette Wing (2006, Communications of the ACM) named the four pillars: decomposition, pattern recognition, abstraction, algorithms. Papert (1980, Mindstorms) called programming "a place where you meet your own thinking explicitly" — code forces vague ideas into precision. Scherer, Siddiq & Sánchez Viveros (2019, Educational Research Review) meta-analysis of 105 studies: near transfer to other programming tasks is robust; far transfer to math or language is limited. Don't sell it as a universal cure, but its value as direct thinking training stands.

Using AI = writing a prompt = decomposing a vague need into precise machine-executable instructions. That is the same muscle as programming. A child who can't decompose only gets surface answers from AI — the gap between "write an essay" and "write an essay on X, from Y angle, in Z words, avoiding W" is enormous.

① Treating coding as a hobby-class KPI (Scratch level, Python level) — trading depth for headcount. ② Caring more about which language is being taught than whether the child decomposed anything today. ③ Expecting coding to broadly lift other subjects — research doesn't support it; treat it as standalone thinking training and the goal becomes clear. For Mom herself: after years of technical work, the most valuable thing you carry isn't a language — it's the muscle memory of decomposition and debugging. That's exactly what you can pass on and what AI can't take.